Estimation of superresolution images using causal networks: the one-dimensional case

Estimating superresolution models from low-resolution sensor data is of great interest for many applications in image processing and computer vision. However, in general the estimation of super-resolution models is difficult due to the computational complexity of existing methods. In this paper, we present an approach to estimating super-resolution world models using stochastic causal networks. The basic elements of our approach include the use of stochastic sensor models, the computation of spatial representations based on random field models, and the development of stochastic estimation procedures to compute these world models from sensor observations. The approach requires only polynomial effort for computing both single cell marginals under arbitrary observations and maximum a posteriori probability (MAP) solutions. We also present approximate methods that further decrease the computational effort for model updating using multiple observations per sensor